Conclusion: WHAT HUMAN BEINGS CAN’T DO!

To the limitation that adversaries of artificial intelligence ascribe to the computer, it could be put forward those of any human being or of the human kind in general. The philosophical standpoint consists in considering these limitations as ontologically linked to the machine. But we already could notice that some of those were common to both supports of “inferences capacities.” This philosophical opposition between human being and machine looks like a simple updating of the traditional opposition between divinity and humanity, in which human being has become the new divinity. The novelty consists in inverting Cartesianism: the machine is too abstract, and the richness of human being is his materiality.

Computer scientist’s understanding

One of the paradoxes of the AI critique concerns understanding and its limit. This question reminds the opposition of notions as comprehension/explanation, beloved of sociology of knowledge. To shorten endless debates, we could define understanding as the sharing of representations, while explanation is equivalent to the reduction of a new concept to established notions (one of the origin of this confusion can come from the fact that in logic, or mathematics, to define in comprehension amount to explanation); we can also say that we explain to others, while we understand ourselves. This justify we do not need to analyze what we understand (from which the philosophical intuition), while explanation is itself an analysis, with the communicative constraint of control by others.

It can be admitted that confusion is possible. But one seems to demand human being when partisans of AI a lack of this ambiguity considered furthermore as their specificity – and as a limitation of understanding for computers. It is shown, in partisans of jargon, by a classic purists’ demand:

“Speaker A say ‘snow is white’ and B can point to the murky grey polluted stuff at their feet. A replies ‘I meant pure snow,’ and B responds ‘You didn’t say so, and anyway no snow is absolutely pure’.” (Winograd & Flores, p. 55).

In the AI framework, we could characterize this dialogue by the principle of typicality characterizing entities by their usual or average characteristics (it is elsewhere the only reason to pronounce some obviousness like “the snow is white”). This kind of anti-communicational semantic purism contradicts the original intention of AI adversaries. We can see very well here that the academic bias leads to make the human being talk like a computer, as it is imagined, a bit stupid. One can even recognize there, not the “clumsiness,” but the formalistic pedantry of the caricature of the scientific egg-head who would state there is water in the refrigerator because there is some in the “eggplants cells” (idem, p. 98).

In the Minsky’s use of the term understand about the Bobrow’s STUDENT program, it could be easy to admit, precisely in a human way, the exaggeration possibility or the relativity of the term to understand. One can very well imagine that the program understands simple sentences, or specialized requests, without understanding subtleties. We can elsewhere observe that Weizenbaum, one of the main parties involved, do not agree with Dreyfus’ maximalist opinion. Elsewhere, he perfectly clarifies the memory systems functions. His program, ELIZA, was able to memorize answers, and to take account of them in its next replies. Despite his reservations, this author finally admit, regretfully, that: “Still it could have been said to have ‘understood’ anything in only the weakest possible sense” (p. 189). So, there is no problem left then, because precisely, the central issue in understanding, is the weakest sense, and especially the pragmatic, operative, consequences. And these results are still very superior to some human beings ones. Even it is possible that intention partisans can always claim to understand more.

If the optimistic predictions or the extrapolations of AI partisans can – with just reason – be taken cautiously, the phenomenological hypothesis, by mentioning the non-finiteness of our thinks, actually describes very precisely this human intellectual program allowing to extrapolate, or to understand extrapolations, projects. Criticizing them for the sake of a principle asserting this very comprehension amounts to dogmatic insincerity, refusing disturbing conclusions. The fact that computer scientists master this understanding better than philosophers can elsewhere leave to imagine that it is automatable. It is explicit anyway in AI partisans like Jacques Pitrat (in Dreyfus’ French edition, p. 434). As he clearly notices that too optimistic predictions are the fate of all research. He even ventures the hypothesis it allows not feeling outdone by difficulties. Conversely, the philosophical questioning seems to consist in scaring the poor human beings by the displaying of obstacles they have to face up. Like in the prehistoric spatial conquest case (see Extrapolation and First Steps), Pitrat is more understanding than Dreyfus does. According to the Surcouf’s principle, those talking the most about understanding are those who seem most need it. While the computer scientists’ mistakes, if any, are more human than the judgement demanding a robotic perfection, while denying its possibility. These numerous contradictions make us doubtful of some philosophers’ mental health or analysis ability. When Pascal simply stated:

“In general we don’t use language, according to strict rules – it hasn’t been taught us by means of strict rules either.” [NOTE 177] (in Dreyfus, p. 203).

It opposed, actually, common sense to logic – even if he was wrong. The Dreyfus’ preference seems nevertheless to go to the Wittgenstein’s gloss, against his phenomenological convictions. We can observe that it does concern a regression:

“But Wittgenstein did not base his argument against the claim that language was a calculus solely on a phenomenological description of the nonrulelike use of language. […] He assumes […] that all nonarbitrary behavior must be rulelike, and then reduces this assumption to absurdity by asking for the rules which we use in applying the rules, an so forth.” (Dreyfus, p. 203).

The so-called reductio ad absurdum is here a contradiction (absurd demonstration and not reductio ad absurdum): consisting in demanding an automation of the discourse for the sake of the theory denying this possibility. Rules in question are nevertheless those allowing holding this kind of reasoning. It is precisely this formalism, in the maniac standardizing meaning, which seems to be an Anglo-Saxon pathology. The validity of Dreyfus’ reservations could be limited to this cultural context.

Conversely, one can consider formalization of natural language as always realized, at a given competence level! One can even to consider it is also the case for computer languages themselves, and to each of their evolution level. Each of them is an intermediate dialect between human being and the machine, since the computer always works in binary encoding, and that each programming language is translatable in the inferior level [NOTE 178].

Anyway, the point of the AI debate is to represent a formalization of questions raised by some members of the university community. This work has therefore at least a documentary value, even anthropological one, as the comments below remind it:

“Within many generations, when AI will have reached its purposes, our successors won’t understand the interest of our papers; they will think we were only telling platitudes, that our papers were only a mass of common places. By reading works like this [Dreyfus’] one, they will reckon these common places were not so obvious, since some competent, cultivated minds, sincerely seeking to clarify them, went to a lot of trouble to admit them.” (Jacques Pitrat, “Discussion,” in Dreyfus’ French edition, p. 439).

But Pitrat is too modest yet: when AI will have reached its purposes, it will be also able to define conditions of a passage from one stage to another, or to clarify what is wrong in a reasoning. If these papers (pro and cons) are not ridiculous, AI will formalize the interpretation itself, as formalization alone is an explanation.

Automated critical method

We can elsewhere reduce this comprehension problem to the acceptance of whatever divergent problematics. It is clearly the case of any criticism, for it is often difficult to appraise the constructive value of criticizing or contesting against us. Elsewhere, their own authors consider themselves as opponents. But then the expertise elicitation also demands, of course, salvaging the critical tradition. From this point of view, the taking over refusal, beloved of leftists, is simply an attempt of adversary inhibition.

Some criticizing are also discredited by the lack of respect of the admitted methodological forms, as we can have seen in the Dreyfus’ case, and in others ones. Thus, critiques like those of our quartet (incidentally a quintet) could be used as positive critiques. But this expertise elicitation implies a constructive tour and an open minded one. It is not our authors’ subjective intention ones, expressed on the imprecation, taboo, and anathema mode. But it is actually possible to neutralize intentions, contrary to these authors’ ideas asserting incommensurability of subjectivity.

If we want to escape the suspicion strategy, without falling in otherworldliness, the ideal way is an automated procedure allowing to take in account the risk of rhetorical abuses and to provide an appropriated answer. An automatic processing has some impersonality and exhaustiveness advantages, especially against insincerity or naivete. It can be considered as an arbitrator (like a dictionary for word games, or the fact checkers for the newspaper papers statements [NOTE 179]). Used at the source, it could constitute a first reading and could allow the writer a better acceptance of criticizing or a treatment of contradictions. Incidentally, it allows to eliminate the plagiarism possibility, and to assess the authentic inventions. And at least, it allows a locating of common places in epistemological works (like Marxist, constructivist, relativists, or phenomenological clichés). We can observe often their persistence, whose ritual routine seems to act as thought for the suckers.

The fundamental principle of the AI challenge is the simulation of the normal behavior of intellectual activity. This justifies the analysis of common sense or natural language, and the point of the idea of thought as calculation is to allow an automation of critique itself. The very critical method used here is elsewhere a step in the automation of information processing included in the quoted texts, or in those they quote themselves.

A text can be considered as a structured database of a particular kind. The specification is extracting knowledge or arguments in order to allow a more convenient look up. We can represent the problem, for instance, by simulating what (always) happens when we read a document. The first phase is a more or less elaborate indexing, since the librarian cataloging until the analysis of each minimal meaning unit. In a reading note, this unit is a quotation or something like a summary. Analysis is not inevitably ultimate, even for a human being, which allows rereading.

In a reading modeling outlook, it can be admitted that we compare elementary phrases with a knowledge base expressible in an other phrases form. They include personal beliefs, external documentary references, a reference to the other phrases of the same text or of the corpus in question, and observation phrases by confrontation to the real. This is of course an obviousness (which negation implicate contradiction), since we cannot oppose to all or a part of this thesis without confirming it.

We can consider the whole debate as the seeking for maintaining an elementary phrase consistency, like in the expert systems knowledge bases. We could have noticed too that inter-contradictions between AI opponents could make them convince one another of their standpoint falseness. Indeed, these contradictions, or one opponent’s internal contradictions, do not constitute a logical proof of the opposite (to say both a and non-a, neither prove a, nor non-a). But it can be admitted that the opposition to the existence or to the possibility of a field like AI demands a minimum of coherence. At least in order not to repeat arguments, or established demonstrations.

We can categorize this coherence for each criterion of validity: subjective criterion (Weizenbaum says X), formal criterion (Searle says X/Searle says non-X), intersubjective criterion (Dreyfus says X/Piaget says non-X), objective criterion (Winograd and Flores say X/observation says non-X), etc. A simple classification by topic already produces a result identifying trends in epistemological arguments. The objective/subjective opposition is equivalent to phenomenology & constructivism Vs empiricism one; the objective/formal opposition is equivalent to the rationalism Vs empiricism one; the subjective/intersubjective opposition is equivalent to the individualism Vs collectivism ones, romantic or Marxist for instance; etc.

One point of this critical standpoint is to avoid sampling problems since this method can work in any circumstances, as soon as the current knowledge. The other point is to be able to formalize any knowledge levels: not to understand anything is already an indication – at least non-matching with the current knowledge – and each comprehension stage is exploitable. This allows admitting a partial criticizing as legitimate. It is thus possible to suppose that the competence nature is merely cumulative, and progressive, like in reading, or in foreign language learning. The last point is to represent the comprehension mistakes, the prejudices making, rumors and legends themselves, by processes identical to exact knowledge ones. This was a condition raised by sociologists of knowledge like David Bloor, for human sciences epistemology.

The only reservations opposed to this automation remains subjectivity. It has become trite to hear that reading is a personal relationship with a text. But this cliché does seem being only a simple academic laicization of the protestant approach of the relation to the Bible! The entire question becomes therefore an objectivation of cognitive phenomena and their simulation by a program or an autonomous system. The result can become a cognitive identity, reduced or not.

We can see although the problem isn’t really computerized comprehension, since it is perfectly possible to undertake the encoding manually. The automatic comprehension is simply, like ever, equivalent to the automation of these manual-encoding processes. It would be possible too of considering indexation as only partial. This automatic indexation becomes a reasonable objective of partial comprehension, which would only process summaries. The overall comprehension becomes then a simple progressive improvement of these competences. Differences with the individual acquisition of knowledge, or its specific formalization, only rests on the fact we have to deal with the mass of already accumulated knowledge.

Cognitive pathology and therapeutic philosophy

It seems possible to assert that essentialist deliriums are equivalent to a pathological state. Indeed, the term of pathology does risk to constitute a labeling, due to external characterization, and to social condemnation habits. But, we can also consider that epistemological obstacles, like those studied by Bachelard, are at least incomplete stages the subject has to overcome. Elsewhere, these epistemological obstacles are only a resumption of Bacon’s idola. The question of the acknowledgement of persons overcoming these obstacles doesn’t solve the problem of knowing how they did it. This can therefore be stated pedagogically speaking. A corrective solution can look like the reframing strategy, put forward by the Palo Alto school psychologists, whose purpose, except possibly some tricks, is simply to explain the changing of beliefs, or to explain this epistemological obstacles overcoming.

A good example of block is Searle, in front of answers to hiss famous Chinese parable (see Chinese semantic). The philosopher seems not to understand what is explained to him, because he certainly doesn’t understand what an explanation is. By his function, the philosopher ought to produce canonical knowledge, by exactly referencing it. But he seems incapable to assimilate a contradiction that would reorganize his knowledge basis – what is after all a rather common attitude.

But there isn’t any more reason to consider only dramatic situations of psychological blockage. As much these obstacles, a posteriori obvious, are legitimate in time. Contrary to what Bachelard seemed to consider, the contemporaries always have good reasons to believe what they believe. More, protagonists’ categories include too, pretty obviously, the scientific concepts too, what Bachelard also seemed to neglect. But then science precisely is the means to improve the formalization of usual or traditional concepts. It agrees again with the hypothesis of continuity of knowledge progress.

Finally, the fact that a common sense mistake can be formalized in logical terms (cf. Boudon, L’art de se persuader des idées douteuses, fragiles ou fausses [The Art of Self-Persuasion: The Social Explanation of False Beliefs]), is the least it could do. This does not mean however that a pre-logic mentality does not exist. But rather those logical operations are not innate or not innately finished anyway. As Boudon said itself, adults make many logical mistakes, which therefore allow to distinguish stages, and to attribute a value to learning. It can be either spontaneous, by simple maturation; or simply pedagogical; or a learning, including a subject’s active participation. It is therefore actually possible to speak, not of pre-logic mentality, but of logical stages characterizing the different mentalities, ideologies, cultures – individual or collective. The fact that we meet a coexistence of these stages in a same culture, or even in an individual alone, is not contradictory. Evolutionism is not even refuted, at the very most it is qualified by the necessity to undertake distribution studies.

Basically, one of the problems having motivated this work was to know how intellectuals could produce howlers being considered as thoughts. In the 20th century course, experience has shown that human thought could be compromised in any ideologies. Labeling oneself humanity cannot be granted against what machine is supposed of being characterized. This problem is equivalent to the reconstitution of mechanisms of these ideologies, or why not to the development of a program simulating these ideological behaviors. We already have seen (in Phenomenological Robotics) that it could be possible, at least partially. And this does answer to the philosophical opposition to thought and reasoning automation. Possibly, it could be admitted that a legitimate anxiety had been shown this way, but by evacuating the responsibility on an artificial scapegoat.

A first contradiction of philosophers and phenomenology focuses on the persistence, in their own discourse, of philosophical categories they are opposed to. But it is even more contradictory to notice that their standpoint, when it legitimately, or not contradictory, opposes to traditional philosophy, is precisely compatible with artificial intelligence. Contrary to their intentions, many phenomenological arguments make possible the AI project. For the novelty of AI is to represent a synthesis between different standpoints, and to solve many paradoxes and enigmas of the philosophical tradition. At least, it could be an intellectual productivity tool [NOTE 180]. But it can nevertheless be said intelligent, either by the mechanisms implemented, or by its object. For whom practicing data processing, the mere paragraphs manipulation into word processors, allows to acknowledge discourse making.

Is coming out for defense of AI, as one of its adversaries did propose, thinking to be a computer. The reader will judge if what he or she just have read is the expression of an individuality. The amateur of paradox will notice that, in this event, it could provide an argument against automaticity. However, the search for repeatability remains a scientific and pedagogical objective, which has to be distinguished, I hope it can be seen, from conformism.

Notes